DISCUSSION PAPER SERIES

IZA DP No. 1345

Prize and Risk-Taking Strategy in Tournaments:

Evidence from Professional Poker Players

Jungmin Lee

October 2004

Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
 Prize and Risk-Taking Strategy
in Tournaments: Evidence from
Professional Poker Players

Jungmin Lee
University of Arkansas-Fayetteville
and IZA Bonn

Discussion Paper No. 1345

October 2004

IZA

P.O. Box 7240

53072 Bonn
Germany

Phone: +49-228-3894-0
Fax: +49-228-3894-180
Email: iza@iza.org

Any opinions expressed here are those of the author(s) and not those of the institute. Research
disseminated by IZA may include views on policy, but the institute itself takes no institutional policy
positions.

The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
company supported by Deutsche Post World Net. The center is associated with the University of Bonn
and offers a stimulating research environment through its research networks, research support, and
visitors and doctoral programs. IZA engages in (i) original and internationally competitive research in
all fields of labor economics, (ii) development of policy concepts, and (iii) dissemination of research
results and concepts to the interested public.

IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.
IZA Discussion Paper No. 1345
October 2004

ABSTRACT

Prize and Risk-Taking Strategy in Tournaments:

Evidence from Professional Poker Players
This study examines whether people optimally respond to prize incentives for risk taking in
tournaments. I exploit the television game show World Poker Tour as a natural experiment.
The results show that professional players strategically choose the degree of risk taking
depending on the incentives implied by the prize structure they face. I find that they are more
sensitive to losses than to gains.

JEL Classification: M5, D8

Keywords: risk, tournament, poker

Jungmin Lee
Department of Economics
University of Arkansas-Fayetteville
Austin, TX 78712
USA
Email: JLee@walton.uark.edu
1 Introduction

Many firms use rank-order tournaments to motivate, assess, and compensate their workers.
Rank-order tournaments are advantageous when it is costly to evaluate individual perfor-
mance in absolute term and when such measure, even if it is affordable, is likely to be
contaminated by common shock (Prendergast 1999). Examples are the competition among
vice presidents for promotion (Main et al. 1993) and rank-based contracts with suppliers
(Knoeber and Thurman 1994). Lazear and Rosen (1982) show that firms can induce the first
best level of effort by using a tournament.
On the other hand risk-taking incentives are intrinsic in rank-order tournaments. Trailers
have an incentive to take high-risk actions during the course of competitions, in particular near
the end, because the prize is the same whether they lose by a little or a lot as long as ranking
does not change. Many examples can be found in business contexts, such as risk taking by
mutual funds and broiler chicken farmers (Chevalier and Ellison 1997, Knoeber and Thurman
1994) as well as in professional sports (Bronars 1987, Becker and Huselid 1992, Bronars
and Oettinger 2001). In sports, risky behaviors are not troublesome, but rather enjoyable.
Spectators enjoy uncertain results and are excited at upsets by underdogs. On the other hand,
in business organizations, risky behaviors and uncertain outcomes could hurt profitability.
They might increase the variability of year-to-year performance, make evaluation and future
planning difficult, and finally destroy investors’ confidence in the market.
In this study I exploit the popular television game show “World Poker Tour” on Travel
Channel in the United States as a natural experiment for examining risk-taking behaviors
in rank-order tournaments. Poker tournaments are suitable for the purpose of this study
since risk taking, such as bluffing and “all-in,” is the most essential component of the game
strategy, while the problem of effort choice is quite trivial. They also provide a unique
opportunity to evaluate risk taking behavior under well-defined rules in the face of high
monetary incentives, which is not affordable in a lab experiment. Furthermore I expect that
unobserved heterogeneity in preferences should be minimal since most players are homogenous
in quality. Lastly it is unlikely that professional players make systematic mistakes in statistical
calculation.
I examine whether professional players strategically choose the degree of risk taking in

2
response to the prize incentives they face. I exploit variation in risk-taking incentives both
within a tournament and across tournaments. The incentives would change for each individual
player within a tournament depending on his or her current position on prize structure. Some
players face relatively large expected gains and small expected losses in prize from a risk-
taking strategy than others. Also the incentives would be different across tournaments since
prize structures, such as total prize and its distribution, significantly differ tournament by
tournament. Poker tournaments should provide a clear view of how people respond to risk-
taking incentives in rank-order tournaments.

2 A Simple Model of Risk Taking Strategy

In a poker game called No Limit Texas Hold’em, each player is given a pair of cards (the
so-called “hole cards”). There are five “community cards” in the end, and the player with
the best seven-card poker wins al the money wagered. As community cards are revealed in
three stages–“Flop” (three cards), “Turn” (one), and “River” (one), players choose whether
to check, fold, or call and how much to raise without knowing other players’ hole cards.
Suppose that there are six players on table. Consider a player i with ci number of chips
at a moment. Subscript i also denotes the current rank. Let wj denote the prize for final
rank j. So wj > wk if j < k. Let pj (ci ) denote the subjective probability of player i that
he will end up with final rank j, where j = 1, 2, ..., 6. It is reasonable to assume that the
probability distribution has a mode at wi , that is pi (ci ) ≥ pj (ci ) for all j 6= i.
For simplicity assume that players choose only whether to call or fold for a certain amount
of bet. In other words my model does not consider how much to bet; players can bet more
chips only by participating more frequently. Suppose that there are n players who calls a
bet, x chips. If player i participates and wins, total chips increase by (n − 1)x, but if he
loses, total chips decrease by x. Of course, x is smaller than or equal to ci because players
cannot bet more chips than what they have. The probability distribution of final ranks will
accordingly change to pj (ci + (n − 1)x) or pj (ci − x), respectively.
First consider a middle-ranked player, that is i 6= 1 or 6.1 Let πi represent his luck (good
1
I assume that there are always six players. In reality it is possible that a player of i 6= 6 is not middle
ranked as some players are eliminated by bankruptcy.

3
or bad hands). That is, it represents the probability of winning. The probability distribution
of final ranks will be that of a “compound lottery” of pj (ci + (n − 1)x) and pj (ci − x) with
weight π. Let pj (ci , x) denote the new probability distribution.2 It is reasonable to assume
that the distribution of pj (ci , x)’s is more spread out than that of pj (ci )’s because betting
always entails some risk taking. Player i participates if and only if:

Σ6j=1 pj (ci , x)U (wj ) ≥ Σ6j=1 pj (ci )U (wj ). (1)

The participation constraint can be rewritten to:

pj (ci , x) − pj (ci )
Σj6=i U (wj ) ≥ U (wi ). (2)
Σk6=i [pk (ci , x) − pk (ci )]

Alternatively,

pj (ci , x) − pj (ci ) pj (ci , x) − pj (ci )

Σj<i U (wi + gij ) + Σj>i U (wi − lij ) ≥ U (wi ),
Σk6=i [pk (ci , x) − pk (ci )] Σk6=i [pk (ci , x) − pk (ci )]

where gij represents the gain in prize and lij the loss in prize when ranking changes from i to
j. Notice that pj (ci , x∗ ) − pj (ci ) > 0 because the distribution is more spread out by betting.3
Therefore the left-hand side of the inequality is sort of a weighted sum of gains and losses.
The participation constraint clearly shows that if other things are equal, the larger g
increases the incentive to take risk. On the other hand, the larger l decreases the incentive
for risk taking. Notice that g’s and l’s are different across individual players depending on
their relative position on the prize structure. In the sample of poker tournaments I will use
later, when we consider only gains and losses from one-rank change, that is gi,i−1 and li,i+1 ,
the average gain is $282,356 for the second rank and $35,772 for the fifth rank. And the
average loss is $151,173 for the second rank and $22,064 for the fifth rank. They also differ
significantly across tournaments because each tournament has different prize structure. It is
therefore testable whether individual players’ risk taking behaviors (betting more frequently
2
I omit π because it does not play a significant role for empirical purpose. I assume that it is statistically
identical for all players. In other words the probability that a player has good or bad hands is same across
individuals by the law of large number.
3
I assume that pj (ci ) = pj (ci , x = 0). That is, the probability distribution does not change much when the
player does not bet. Note that the assumption is not always true because the distribution could change by
changes in other players’ chips even though the player does not participate.

4
and probably more chips) are affected by expected gains and losses they face.
The result here explains why prize structures are usually convex in poker games or sports
events where spectators enjoy uncertainty, suspense, and upsets. The convexity implies that
g’s are larger than l’s regardless of ranking. It increases risk-taking incentives overall.
Risk-taking strategy is relatively trivial for top-ranked and bottom-ranked players. Top-
ranked player face no expected gains from a risky strategy and so should try to lock in
the current ranking. They would participate only when they have good hands. Top-ranked
players’ incentives should be minimal among players. On the other hand, bottom-ranked
players would have relatively stronger incentives for risk taking than others because they
have nothing to lose. For these players the current position is a guaranteed place no matter
what they do.
Lastly, the participation constraint also depends on marginal changes in probability dis-
tribution by risk-taking strategy. For example, if pj (ci , x)−pj (ci ) is small for j < i, it is likely
that the player will not participate. This suggests that when the player follows the leaders
very closely, risk taking strategy does not improve the probability of advancing significantly
and so risk taking is not so attractive. Instead he might prefer more prudential plays to wait
for the best chance. On the other hand, if pj (ci , x) − pj (ci ) is small for j > i, the player is
likely to fold his cards. That implies that if the player leads the followers by relatively large
gaps, it is more attractive to take risk. Even if he loses some chips, it is not likely that the
followers catch up with him. Risk-taking incentives depend on chip spreads among players
as well as prize structure.

3 Data from World Poker Tour

The data I use in this study are about individual players’ performance, ranking, and money
prize in different tournaments. Information at the tournament level, such as total prize and
variance, is collected from the official website of “World Poker Tour” (www.worldpokertour.com).
The data are available for 27 tournaments in season one (2002-2003) and season two (2003-
2004). For these tournaments we can observe predetermined prize structures, six players at
the final table, their initial chips, and final ranking. Each player starts with different number
of starting chips, which are cumulated through preliminary rounds. There is no information

5
on preliminary rounds.4 The final round is an elimination tournament in which one is out of
the table if and only if he or she is bankrupt. All chips will be eventually concentrated to
the final winner’s hands.
Table 1 presents prize structures for 27 tournaments. Total money prize is huge; it is on
average 1.4 million dollars. Total prize varies significantly across tournaments. The biggest
tournament is WPT Championship, 5.4 million dollars. The distribution of prize over ranking
is very unequal and convex. The final winner takes almost a half of total prize. Top three
players take 81 percent of total money. Compared to the last-place prize, the first prize is 11
times larger, the second prize is 5.4 times larger. But the fourth prize is 1.9 times larger, and
the fifth prize is only 1.3 times larger. As with total prize, the distribution is also significantly
different across tournaments.
As mentioned, it is possible to identify individual players’ initial and final rank. The
extent to which players’ ranks change during a tournament is informative in the sense that
there are more changes if more players take high-risk strategies. For the purpose of this
study it would be interesting to examine if there is any systematic relationship between prize
structure and ranking changes. I examine two features of prize structure; dispersion and
total prize. If ranking changed only by luck, they should be independent of rank changes.
By regression I find [Changes in Rank ] = 6.28−1.72×[Prize Dispersion]−0.34×[Total Prize]
where N = 27 and R2 = 0.18.5 The estimates for prize structure are statistically significant at
10% and 1% significance level, respectively. Figure 2 shows that there is a small but negative
relationship between changes in rank and prize dispersion. It also shows the relationship
between changes in rank and total prize. The result for prize dispersion is consistent with the
standard expectation that higher dispersion should decrease changes in rank after controlling
for mean as long as players are risk averse. The result for total prize is consistent with
the experimental finding of Kachelmeier and Shehata (1992) that people become more risk
averse at higher stake. This simple look at the data already confirms that players’ risk-taking
behaviors depend on prize structure.
4
There are 80-150 entrants in a tournament. The buy-in dollars (minimum entry fee) are significantly
different, ranging from 300 to 25,000.
5
I measure changes in rank by number of players whose final rank is different from initial rank. Prize
dispersion is measured by coefficient of variation (standard deviation divided by mean). The unit of total
prize is 1 million dollars.

6
 The more detailed data at the individual-player level can be collected for only 12 tour-
naments from the World Poker Tour DVD collection (season one).6 The data contain more
information since we can actually watch each episode and follow individual players’ perfor-
mance (chips) and ranking within a tournament. I keep record of chip counts whenever they
are shown to viewers. It is not possible to construct the complete history of changes in chips
because chip counts are not displayed on the play-by-play basis. I collect the data on indi-
vidual holding of chips only until there remain at least three players, first, because I focus on
middle-ranked players and, second, because chip counts are not usually available after only
two players remain.
Table 2 shows some descriptive statistics. There are 286 observations for middle-ranked
players and additional 106 observations for bottom-ranked players. Players are categorized
as bottom-ranked players when they are ranked the last among surviving players. The first
thing to be noted in Table 2 is that bottom-ranked players are riskier than middle-ranked
players in terms of the amount of chips they bet. Bottom-ranked players bet 60 percent on
average while middle-ranked players bet about 30 percent of their chips. This is because
bottom-ranked players bet almost same amount of chips as middle-ranked players do even
though they hold only a half of chips. By the convexity of prize structure, the expected gain
in prize by one-rank advancement is much larger than the expected loss by one-rank retreat.
The expected gain in prize by advancing a rank is much smaller in size for bottom-ranked
players ($27,022) than for middle-ranked players ($100,192). However the expected loss by a
risk-taking strategy is zero for bottom-ranked players, which explains higher risk taking by
bottom-ranked players.

4 Empirical Analysis

4.1 Measure of Risk Taking

The model shows that players take risk by participating and betting more frequently when
there exists a stronger incentive for risk taking. So I measure the degree of risk taking for an
individual player’s strategy by the absolute value of variation in chips over the tournament.
This is in the same spirit of Knoeber and Thurman (1994) in which they measure the degree
6
I dropped one tournament, Aruba Poker Classic, since its game format is very different from the others.

7
of risk taking in production by a measure of variability in output.7 Specifically my measure
of risk taking is ∆cink = |c·,n,k − ci,n−1,k | where | · | means absolute value. ci,n,k represents
the amount of chips of the i-th ranked player at the n-th chip count display in tournament k.
It measures the change in chips between the (n-1)-th and n-th chip count. Only the absolute
value of chip change is considered because we are interested in the degree of risk taking itself,
not actual outcomes (success or failure) of risk-taking strategy.
It should be kept in mind that ∆cink is likely to underestimate a player’s actual willingness
to take risk. Poker game is an interactive game. A player cannot raise money unless at least
one other player call his bet. The optimal bet is somewhere between the amount of chips a
player likes to bet and the amount of bet he thinks other players will call.
The chip change in absolute value, ∆cink , might be mechanically positively correlated
with the amount of chips that the player initially holds at the (n-1)-th chip count. In other
words ∆cink could simply reflect the fact that players with more chips bet more and have
more changes in absolute value. This suggests that we need to control for the amount of
chips at the previous chip count, ci,n−1,k .8 For robustness check I also experiment with an
alternative variable (%∆cink ), which measures the percentage change in chips between two
chip count displays.

4.2 Broadcasting Bias

I can observe individual players’ chips only when a tabulated report of individual chip counts
is shown to viewers. The editors of the TV show determine how many times and when to
show chip counts. As a result, the measure of ∆cink does not represent actual fluctuation in
individual holding of chips and actual degree of risk taking. My findings in this paper should
be taken with a grain of salt in this aspect.
However it seems reasonable to assume that ∆cink approximates quite well actual changes
in chips. I assume that the editors of the TV show would want to maximize the excitement of
viewers. Then they should inform viewers of changes in chips whenever there are significant
7
However the variability in output might not represent risk taking strategy in the case of agricultural
production. The variability could reflect mismanagement or technological defects. If so, the correlation
between output variability and ranking could reflect simply heterogeneous managerial capabilities or differences
in technology.
8
Instead we could control for total number of chips. But it does not change the results.

8
changes in chip spread or ranking. It is not likely that chip counts are shown when chips
change little. Chip counts are almost always shown when a player is bankrupt and eliminated
from the tournament. So we can identify at least to which players the chips that the eliminated
player held are redistributed.
Also it should be noted that changes in chips between chip counts are likely to underesti-
mate actual degree of risk taking because the measure of ∆cink omits some minor fluctuations.
If so, my estimate of the incentive effect on risk taking strategy would be a lower bound.

4.3 Specification and Results

The basic estimation equation is:

∆cink = α1 gink + α2 link + β1 σ − (cink ) + β2 σ + (cink ) + Xink γ + ²ink , (3)

where n increases within a tournament and n = 1, 2, . . . , Nk . Explanatory variables g and

l are crucial for the purpose of this paper. gi is the gain in money prize by advancing one
rank.9 It is the prize gap between wi and wi−1 . And li is the marginal loss in money prize
by one rank. It is the prize gap between wi and wi+1 . Formally I specify that gi = wi−1 − wi
and li = wi − wi+1 . The model predicts that α1 is positive and α2 is negative.
Individuals’ risk incentives depend not only on neighboring prizes, but generally the whole
prize structure, {w1 , ..., wi , ..., w6 }, as shown in the model. Including the detailed information
on risk incentives, such as wi−2 − wi and wi − wi+2 , reduces the sample size. Thus gi is a
proxy for incentives from higher prizes and li represents incentives from lower prizes.
One might wonder whether risk-taking incentives could change between chip counts be-
cause individual plays between chip counts are omitted and unobservable. Then g and l
constructed based on the ranking at the (n-1)-th chip count would not measure risk-taking
incentives that can explain changes in chips between chip counts. Table 3 addresses the
issue; it shows that ranking changes little between chip counts, even though chips change
a lot as seen in Table 2. About 50 percent of players do not change ranking at all, and
35 percent change only one rank. It is reasonable to assume that risk-taking incentives are
rather constant between chip counts.
9
For simplicity I will omit subscripts unless necessary.

9
 σ − (cink ) and σ + (cink ) represent chip spreads from the nearest leader and follower, re-
spectively. Formally, σ − (cink ) = (ci−1 − ci ) and σ − (cink ) = (ci − ci+1 ). For example, as a
player leads the nearest follower by a relatively large gap, it is likely that his current ranking
is guaranteed. Then, risk taking is more attractive. On the other hand, as he is trailed by
the nearest leader by a small gap, it is likely that he can catch up by normal plays. Risk
taking is then not necessary. The coefficients for both variables are expected to be positive.
However, unlike g and l, chip spreads might change much between chip count displays. So the
measures of chip spreads at the (n-1)-th display could only weakly represent actual concerns
about leaders and followers.
Xink is a vector of a constant and control variables, including ci,n−1,k , wnk , Nk , and n. The
minimum prize (wnk ) at the n-th chip count is included. Note that it is not constant within
a tournament because it increases as players are eliminated. It is a guaranteed amount of
prize at the moment of the tournament. The coefficient is therefore expected to be positive.10
I also control for the number of chip counts in the tournament (Nk ) because I expect that
more frequent chip counts should capture more fluctuations in chips. The maximum number
of chip counts differ across tournaments by the editorial decision. The smallest Nk is 3 and
the largest one is 16 in the sample. n is included since variation in chips would be higher as
a tournament approaches the end and smaller number of players remain.
Lastly ²ink is the error term. I assume it is independent of explanatory variables, but
allow for correlation across individual players for a pair of n and k. Since poker game is
a zero-sum game, there are always winners and losers. Even if we consider the absolute
values of chip changes, they are likely to be negatively correlated among players.11 I correct
standard errors by clustering for a pair of n and k.
Table 4 shows the results for the sample of middle-ranked players. As expected, the
degree of risk taking depends on the incentives that individual players face. An increase in
the gain in prize increases risk taking. An decrease in the loss also increases risk taking. The
average holding of chips is 242,760. The estimates in column (4) suggest that an average
player with 242,760 chips would put on risk about 6.1 percent of chips in response to one
10
It should be kept in mind that wnk is strongly and positively correlated with total prize. The correlation
coefficient is over 0.9. The coefficient for minimum prize could capture any effect of total prize.
11
Consider that two players who win some chips from others play each other. If a player wins, then his
chips in absolute vale increase, while the loser’s chips in absolute value decrease.

10
standard deviation increase in g ($101,059). On the other hand, he would reduce his bet by
about 14.2 percent in response to one standard deviation increase in l ($48,455). The finding
that players are more sensitive to losses than to gains is consistent with the hypothesis of loss
aversion (Kahneman and Tversky 1979). The hypothesis that |α1 | = |α2 | is rejected except
column (1).
The results show that the more chips a player holds, the more chips he bets. An extra
chip increases the absolute value of chip change by about 0.1, much less than one chip.12 As
expected, I find that as a player is trailed by the nearest leader by a larger gap or as he leads
the nearest follower by a larger gap, he would bet more. The effect of chip spread from the
nearest follower is however statistically and economically insignificant. This result suggests
that risk taking is more dependent on chip spread from the leader.
In column (5) I include players’ age and its squared term. Mostly some background
information about players, such as age, career, and residence, are introduced for viewers in
the beginning of show. But I cannot identify age for some players, which reduces the sample
size. Figure 3 is a motivation, which suggests a U-shaped relationship between age and risk
taking. Below average age 46, as players are younger, they are more aggressive; their chips
fluctuate more in absolute value. For older players the relationship is not significant. I find
the same results about age from regression in column (5).
Other results are also consistent with expectation. The minimum prize significantly in-
creases the absolute value of chip change. The more frequently chip counts are displayed to
viewers, the more variation in chips players have. Lastly, as tournament approaches the end,
players bet more chips and the absolute value of chip fluctuation increases significantly.
Table 5 presents the results for the sample of middle-ranked and bottom-ranked players.
For bottom-ranked players l is zero since there is nothing to lose. The chip spread from the
nearest follower is not available for bottom-ranked players, so I have to interact (σ + (c)) with
the indicator for middle-ranked players (Middle Rank ). The results are overall very similar to
those in Table 4. The effect of l is larger in absolute value. This implies that bottom-ranked
players would take more risk because they have nothing to lose in rank-order tournaments.
Again the hypothesis that |α1 | = |α2 | is strongly rejected across the board.13
12
I include the quadratic term of cn−1 for the case of nonlinearity, but find it insignificant.
13
I also run regressions with interacted terms between risk-taking incentives and chip spreads because, for

11
 Table 6 and 7 presents the results for different specifications for robustness check. In
Table 6 I use percentage change in chips (%∆cink ) as an alternative dependent variable. The
results are virtually equivalent to the previous ones. For middle-ranked players one standard
deviation increase in g would increase the absolute value of percentage change in chips by
2 to 5 percent. On the other hand, one standard deviation increase in l should decrease it
by 8.6 to 11.4 percent. As with the previous results, the effects are significantly asymmetric
in size. The results do not change qualitatively when we include bottom-ranked players.
Most of other variables except Nk and n become statistically insignificant. Table 7 shows the
results when some additional control variables are included. I include rank-specific dummy
variables for the case that there is any direct effect of ranking. I also include tournament
characteristics of prize structure (total prize and prize dispersion). The results do not change
much. Additional variables are not statistically significant.

5 Conclusion

This study examines whether people optimally respond to risk-taking incentives in rank-order
tournaments at high stake. I exploit professional poker tournament as a natural experiment.
I find that professional players choose the degree of risk taking depending on monetary
incentives, i.e. expected gains and losses implied from their relative position on the prize
structure. Holding other things equal, a larger expected gain or smaller expected loss would
strengthen the incentive for risk taking. I also find that the effects of expected gains and
losses are highly asymmetric. Players are significantly more responsive to expected losses
that gains.

References

[1] Becker, Brian E. and Huselid, Mark A. “The Incentive Effects of Tournament Compen-
sation Systems.” Administrative Science Quarterly, 1992, 37(2), pp. 336-350.

[2] Bronars, Stephen. “Risk Taking Behavior in Tournament.” mimeo, 1987.

example, players could be more responsive to the expected gains from risk-taking strategies when it is less
likely that they are caught up with by followers. However I find that the interacted terms are not significant
both economically and statistically while they do not change other estimates qualitatively.

12
[3] Bronars, Stephen and Oettinger, Gerald. “Effort, Risk-Taking, and Participation in Tour-
naments: Evidence from Professional Golf.” mimeo, Department of Economics, University
of Texas at Austin, 2001.

[4] Chevalier, Judith and Ellison, Glenn. “Risk Taking by Mutual Funds as a Response to
Incentives.” Journal of Political Economy, 1997, 105(6), pp. 1167-1200.

[5] Hvide, Hans K. “Tournament Rewards and Risk Taking.” Journal of Labor Economics,
2002, 20(4), pp. 877-898.

[6] Kachelmeier, Steven J. and Shehata, Mohamed. “Examining Risk Preferences under
High Monetary Incentives: Experimental Evidence from the People’s Republic of China.”
American Economic Review, 1992, 82(5), pp. 1120-1141.

[7] Kahneman, Daniel and Tversky, Amos. “Prospect Theory: An Analysis of Decision under
Risk.” Econometrica, 1979, 47(2), pp. 263-291.

[8] Knoeber, Charles R. and Thurman, Walter N. “Testing the Theory of Tournaments: An
Empirical Analysis of Broiler Production.” Journal of Labor Economics, 1994, 12(4), pp.
155-179.

[9] Lazear, Edward P. and Rosen, Sherwin. “Rank Order Tournaments as Optimum Labor
Contracts.” Journal of Political Economy, 1981, 89, pp. 841-864.

[10] Main, Brian G. M., O’Reilly, Charles A., and Wade, James. “Top Executive Pay:
Tournament or Teamwork?” Journal of Labor Economics, 1993, 4, pp. 606-628.

[11] Prendergast, Canice. “The Provision of Incentives in Firms.” Journal of Economic

Literature, 1999, 37(1), pp. 7-63.

[12] Rosen, Sherwin. “Prizes and Incentives in Elimination Tournaments.” American Eco-
nomic Review, 1986, 76(4), pp. 701-715.

13
 Table 1: Prize Structure for 27 Tournaments1
Mean Median SD Min Max
First Prize (w1 ) 606,151 500,000 502,048 100,000 2,278,356
Second Prize (w2 ) 323,795 278,240 291,146 38,000 1,372,223
Third Prize (w3 ) 172,622 139,120 153,198 18,000 706,903
Fourth Prize (w4 ) 118,725 83,472 106,854 10,000 457,406
Fifth Prize (w5 ) 82,953 62,604 73,105 7,000 332,660
Last Prize (w6 ) 60,889 46,715 51,645 6,000 232,862
Total Prize 1,365,135 1,091,430 1,165,011 183,000 5,380,410
1
Prizes are denominated in the US dollars.

14
 Table 2: Descriptive Statistics of Crucial Variables1
Variable Mean SD Min Max
Middle-Ranked Players (N = 286)
Absolute Variation in Chips ∆c 62,235 92,881 0 798,000
Current Chips cn−1 242,760 204,839 26,000 1,239,000
Percentage Variation in Chips %∆c 30.0% 33.3% 0% 173.1%
Chip Spread from Nearest Leader σ − (c) 156,997 226,275 0 1,813,000
Chip Spread from Nearest Follower σ + (c) 73,200 79,653 0 511,000
Gain in Prize by One-Rank Change g 100,192 101,059 3,000 505,261
Loss in Prize by One-Rank Change l 47,108 48,455 1,000 253,312
Bottom-Ranked Players (N = 106)
Absolute Variation in Chips ∆c 60,890 79,607 0 391,000
Current Chips cn−1 111,716 107,271 5,000 666,000
Percentage Variation in Chips %∆c 60.3 55.2 0 246.9
Chip Spread from Nearest Leader σ − (c) 65,829 72,081 500 371,000
Chip Spread from Nearest Follower σ + (c) – – – –
Gain in Prize by One-Rank Change g 27,022 31,464 1,000 153,135
Loss in Prize by One-Rank Change l 0 0 0 0
1
Prizes are denominated in the US dollars.

15
Table 3: Changes in Rank between Chip Count Displays

Middle-Ranked Players
Rank Change Frequency Percent (%)
0 140 48.95
1 115 40.21
2 24 8.39
3 6 2.10
4 1 0.35

All Players except Top Rank

Rank Change Frequency Percent (%)
0 204 52.04
1 140 35.71
2 35 8.93
3 12 3.06
4 1 0.26

16
 Table 4: Middle-Ranked Players1
(1) (2) (3) (4) (5)
gi .164 .241 .192 .178 .181
(.076) (.076) (.074) (.075) (.096)
li -.396 -.870 -.792 -.797 -.746
(.209) (.286) (.270) (.267) (.246)
cn−1 .184 .116 .113 .118 .070
(.060) (.059) (.060) (.057) (.058)
σ − (c) .119 .117 .101 .119
(.054) (.055) (.054) (.058)
σ + (c) -.002 .010 -.038 .010
(.108) (.107) (.105) (.101)
w .787 .673 .801 .722
(.443) (.408) (.413) (.336)
Nk -3439 -5795 -5301
(1688) (1697) (2371)
n 5837 6269
(1775) (1940)
Age -5825
(2760)
Age Squared 62.5
(35.3)
Constant 19724 4759 44762 43457 163722
(8992) (14095) (20004) (19389) (53561)
R2 = .153 .213 .228 .257 .280
F-test for |α1 | = |α2 | [.175] [.008] [.009] [.001] [.004]
Number of Observations 286 255
1
Robust standard errors in parentheses are calculated by clustering. p-values are in
brackets.

17
 Table 5: Middle-Ranked and Bottom-Ranked Players1
(1) (2) (3) (4) (5)
gi .232 .228 .201 .135 .141
(.076) (.072) (.074) (.073) (.087)
li -.729 -.964 -.920 -.796 -.793
(.175) (.256) (.249) (.242) (.243)
cn−1 .184 .184 .179 .174 .139
(.054) (.053) (.055) (.051) (.047)
σ − (c) .112 .111 .091 .109
(.051) (.051) (.051) (.055)
σ + (c)×Middle Rank -.081 -.070 -.105 -.079
(.100) (.098) (.096) (.093)
Middle Rank -681 58.5 -9492 -3785
(8588) (8494) (8280) (8841)
w .718 .638 .775 .719
(.363) (.335) (.330) (.320)
Nk -2610 -5270 -5052
(1435) (1457) (1858)
n 5967 6461
(1643) (1765)
Age -4694
(2148)
Age Squared 48.3
(26.0)
Constant 20119 3026 33263 35632 138114
(7088) (11974) (17488) (16575) (43169)
R2 = .189 .248 .257 .292 .307
F-test for |α1 | = |α2 | [.001] [.001] [.001] [.001] [.001]
Number of Observations 392 349
1
Robust standard errors in parentheses are calculated by clustering. p-values are in
brackets. Middle Rank is an indicator of middle-ranked players.

18
 Table 6: Percentage Change in Chips1
(1) (2) (3) (4)
gi .000050 .000020 .000053 .000032
(.000029) (.000033) (.000038) (.000042)
li -.000235 -.000178 -.000242 -.000154
(.000067) (.000078) (.000088) (.000090)
R2 = .041 .121 .120 .161
F-test for |α1 | = |α2 | [.0001] [.0072] [.0012] [.0427]
Number of Observations 286 392
1
Robust standard errors in parentheses are calculated by clustering. p-values are in
brackets. Other variables included are σ − (c) and σ + (c) in column (1), σ − (c), σ + (c)
cn−1 , w, Nk , and n in column (2). Included are σ − (c), σ + (c)×Middle Rank, and
Middle Rank in column (3), σ − (c), σ + (c)×Middle Rank, Middle Rank, cn−1 , w, Nk ,
and n in column (4).

19
 Table 7: Robustness Checks1
Sample gi li Controls |α1 | = |α2 |
Middle .144 -.638 Basic, Dummies for Rank [.077]
(.073) (.313)
Middle .146 -.652 Basic, Dummies for Rank, [.086]
(.083) (.298) Tournament Characteristics
Middle & Bottom .109 -.695 Basic, Dummies for Rank [.010]
(.074) (.260)
Middle & Bottom .092 -.676 Basic, Dummies for Rank, [.014]
(.090) (.253) Tournament Characteristics
1
Robust standard errors in parentheses are calculated by clustering. p-values are in brack-
ets. For the sample of middle-ranked players, basic control variables include cn−1 , σ − (c),
σ + (c), w, Nk , and n. For the sample of middle-ranked and bottom-ranked players, basic
control variables include cn−1 , σ − (c), σ + (c)×Middle Rank, Middle Rank, w, Nk , and n.
Tournament characteristics include total prize and coefficient of variation in prize.

20
 Figure 1: Convex Prize Structures

WPT Prize Structure

2,500,000

2,000,000

1,500,000

1,000,000

500,000

0
Prize6 Prize5 Prize4 Prize3 Prize2 Prize1

21
 Figure 2: Prize Structure and Rank Change

Total Prize and Rank Change

6
5 (Number of Players with Rank Change)
Rank Change
43
2

0 2000000 4000000 6000000

Total Prize

(a) Total Prize and Rank Change

Prize Dispersion and Rank Change

(Number of Players with Rank Change)
6
5
Rank Change
4
3
2

.6 .8 1 1.2 1.4
Coefficient of Variation

(b) Prize Dispersion and Rank Change

22
 Figure 3: Age Profile of Risk Taking
150000
100000
Chip Change
50000 0

20 40 60 80
Age
Note: The predictions are based on polynomial regression with 95% confidence interval.

23